Abstract
In the age of the infodemic, it is crucial to have tools for effectively monitoring the spread of rampant rumors that can quickly go viral, as well as identifying vulnerable users who may be more susceptible to spreading such misinformation. This proactive approach allows for timely preventive measures to be taken, mitigating the negative impact of false information on society. We propose a novel approach to predict viral rumors and vulnerable users using a unified graph neural network model. We pre-train network-based user embeddings and leverage a cross-attention mechanism between users and posts, together with a community-enhanced vulnerability propagation (CVP) method to improve user and propagation graph representations. Furthermore, we employ two multi-task training strategies to mitigate negative transfer effects among tasks in different settings, enhancing the overall performance of our approach. We also construct two datasets with ground-truth annotations on information virality and user vulnerability in rumor and non-rumor events, which are automatically derived from existing rumor detection datasets. Extensive evaluation results of our joint learning model confirm its superiority over strong baselines in all three tasks: rumor detection, virality prediction, and user vulnerability scoring. For instance, compared to the best baselines based on the Weibo dataset, our model makes 3.8% and 3.0% improvements on Accuracy and MacF1 for rumor detection, and reduces mean squared error (MSE) by 23.9% and 16.5% for virality prediction and user vulnerability scoring, respectively. Our findings suggest that our approach effectively captures the correlation between rumor virality and user vulnerability, leveraging this information to improve prediction performance and provide a valuable tool for infodemic surveillance.
Introduction
In the current "infodemic" era, social media platforms are frequently inundated with misinformation. Traditional rumor detection systems typically focus on a binary classification—determining whether a post is a rumor or not. However, this narrow focus is often insufficient for real-world surveillance. To prioritize interventions effectively, stakeholders must understand the potential impact of a rumor and the characteristics of the audience it reaches.
As illustrated in Figure 1, a comprehensive infodemic surveillance system should address three interconnected dimensions: rumor status, virality (how far it will spread), and user vulnerability (who is likely to spread it). By identifying viral rumors early, platforms can prioritize them for manual verification. Simultaneously, identifying vulnerable users allows for targeted "early alerts" to mitigate the spread of false information.
Methodology: A Unified GNN Framework
The authors propose a unified Graph Neural Network (GNN) framework designed to handle these three tasks simultaneously. The architecture is divided into four distinct stages: construction, input embedding, refined embedding, and output layers.
1. Graph Construction
The system constructs two types of graphs: Propagation Networks, which track the flow of specific posts, and User Interaction Networks, which capture the broader social relationships between users. These are merged to create a comprehensive view of how information moves through a social network.
2. Hybrid Embeddings
To represent the data, the model uses a hybrid approach. Post contents are processed using a pre-trained BERT model, while user features are initialized via Self-Supervised Contrastive Learning. This ensures that both the semantic content of the rumors and the historical behavior of the users are captured.
3. Refined Representations and Community Detection
A user-post cross-attention mechanism allows the model to weight user features based on the content they interact with. Furthermore, the model employs DiffPool, a hierarchical pooling method, to identify latent communities. This is based on the insight that users often share behavioral traits within specific social clusters.
4. Community-enhanced Vulnerability Propagation (CVP)
A key innovation is the CVP method, which refines individual user representations by integrating them with their latent community's embedding. This allows group-level signals to inform individual vulnerability scores, acknowledging that users do not act in isolation.
Datasets and Multi-Task Learning
The framework was evaluated using two large-scale datasets: TWITTER (derived from Twitter15 and Twitter16) and WEIBO. Ground-truth labels for virality were based on unique user counts in cascades, while vulnerability was defined as the ratio of rumor-related engagements to total activity.
To manage the potential for "negative transfer" between tasks, the authors utilized a Meta-learning strategy (MT-META). This approach optimizes the model's parameters to find a balanced initialization that performs well across rumor detection, virality prediction, and vulnerability scoring, preventing one task from dominating the learning process.
Key Results and Insights
The joint model demonstrated significant performance gains over specialized baselines:
- Rumor Detection: Achieved a 3.8% accuracy improvement on the Weibo dataset.
- Virality Prediction: Reduced Mean Squared Error (MSE) by 23.9% on Weibo, significantly outperforming models like TCAN.
- User Vulnerability: Reduced MSE by 16.5% on Weibo, surpassing user-centric models like U-BERT.
One of the most critical findings is the "vulnerability bridge": user vulnerability serves as a vital link between rumor status and virality. The model remains effective even in early detection scenarios, maintaining high performance with only 20-40% of the propagation data available.
Conclusion
This research provides a powerful tool for infodemic surveillance by treating rumor detection, virality, and vulnerability as a unified problem. Future work may explore incorporating psychological profiles to deepen the understanding of user susceptibility and adopting ranking-based algorithms to further refine prediction metrics.
Summary updated